More on Analyzing with an Engine

        I’ve mentioned in previous posts the necessity of using more than one engine to analyze a position AND the necessity of, after completing the analysis, skipping to the end of the variation and backtracking through the analysis in order to see if the engine missed anything in its calculations.  This post will show how the technique works.
       In the following position taken from one of my old correspondence games, Black overlooked a mate threat and lost immediately.  I tested three engines, Houdini, Stockfish and FireBird just to see what they came up with.  In most of my analysis the engines were allowed one to three minutes per move to examine the positions.
       After they reached what seemed like a reasonably reliable conclusion on the best move, I pasted the analysis into the game then went to the end of the line and started backing up.  You will notice that the engine’s evaluation of the final position often changed, sometimes significantly, and in the process of backtracking through the line, other improvements were occasionally suggested.  Each of those had to be checked out.  This takes a lot of time and patience to do it thoroughly…something I don’t have.  Well, I have the time but not the patience which may explain why my success at Lechinicher SchachServer has been modest at best.  I’m in too big a hurry to see who wins to spend days and days to properly analyze my games with an engine.  But, that’s what the big boys do and as you will see if you go through the game, it’s why they say you should not simply trust what the engine is telling you.
      Beyond all that though, it also serves as a reminder that when letting an engine analyze your own games, you should not just plug in the game and let it analyze at some give time per move and then believe you have the absolute best analysis.  Everything has to be checked out!
       In this piece of analysis, all three engines immediately saw that Black had to play 23…Rg8 All three were allowed to ponder the position three minutes.  Houdini wanted to play 24.Rxd4 and thought White was better by 0.65 and when skipping to the end of its main line it still considered White’s advantage to be only 0.54.  Backing up a move at a time from the end of the line though showed something quite different. 
       When it was Black’s 35^th move, instead of 35…f6 in the main line, Houdini suggested 35…Rxh4 instead with the position about equal.  Skipping to the end of that line and backtracking through the moves did not reveal any significant changes in the evaluation, so it could be assumed that by playing 23...Rg8 24.Rxd4 Rg5 25.Qf3 Be7 26.Rxd7 Qxd7 27.Re1 Kg8 28.h4 Rb5 29.Re3 Kf8 30.Qf4 Rh5 31.Kg1 Kg7 32.Rg3+ Kf8 33.Qe4 Bd8 34.Rf3 f5 35.Qe5 Rxh4 36.g3 f6 37.Qxf5 Rd4 38.Qxd7 Rxd7 39.Kg2 Kg7 40.Nf4 Kf7 41.Re3 Bc7 42.Rd3 Rxd3 43.Nxd3 a5 44.Kf3 Ke6 45.Ke4 f5+ 46.Kf3 b5 47.Ke3 Bb6+ 48.Ke2 a4 49.Nf4+ Ke5 Black likely would have been able to hold the game.
       FireBird 1.0 x64 followed Houdini’s analysis up to move 27 when it wanted to play 27.Nf4 instead of Houdini’s 27.Re1. Checking out 27.Nf4 with Houdini revealed it wanted Black to meet the move with 27…Qc7 and Firebird’s reply for Black didn’t even make Houdini’s top three recommendations although its numerical evaluation was pretty close to Firebird’s.
       Stockfish 2.0.1 JA 64bit also quickly played 23...Rg8 but wanted to play 24.Rh4 instead of 24.Rxd4 as suggested by the other two engines. Checking Stockfish’s play with the Houdini engine didn’t reveal anything significant but I did notice that Houdini’s evaluations were not quite as optimistic as Stockfish’s.
       SO, the question is, after 23…Rg8, can Black hold the game?  Apparently he can if he follows Houdini’s analysis.